OpenAI names its AI model 'Strawberry' after the task it cannot yet perform: counting the number of 'r' letters in 'strawberry'

In 2024, a simple question circulated widely on social media: “How many letters r are in the word strawberry?” The correct answer is 3. ChatGPT consistently replied 2, and did so with confidence—sparking mockery in the tech community and raising a broader question about how large language models process text.

Why ChatGPT answers 2 instead of 3

The issue is tied to how large language models read language. Unlike humans, who can process characters one by one, large language models break text into tokens—chunks that are often larger than a single character, such as parts of words, prefixes, suffixes, or common syllables.

With GPT-4, the word “strawberry” is not treated as 10 individual letters. Instead, it is represented as 3 tokens: str, aw, and berry. Only str and berry contain the letter “r,” so the model’s token-based representation leads it to count 2.

This behavior is described as a consequence of model architecture—built to understand meaning rather than to parse spelling at the character level—rather than a simple bug that can be patched.

OpenAI’s “Strawberry” project and the Q* link

The story gained additional context in late 2023 and 2024. In November 2023, Sam Altman was fired from OpenAI, and rumors followed about a secret project named Q* that allegedly achieved a breakthrough in mathematical reasoning.

Eight months later, in July 2024, Reuters reported that OpenAI was developing a model with the internal codename Strawberry. The report indicated that the codenames Q* and Strawberry referred to the same effort.

OpenAI later described the “Strawberry” name as an internal declaration: the team was building a final model intended to do what previous models could not—count the number of “r” letters in “strawberry.” When o1 was released in September 2024, OpenAI included the strawberry question as a public challenge in the interface for its older product.

What changed with o1—and why it costs more

According to the account, the improvement was not achieved by changing tokenization. Instead, the model was trained to “think aloud” before answering, using chain-of-thought reasoning. Rather than responding immediately, it was described as counting letters step by step, then checking again—producing the correct answer of 3.

However, this approach has a performance trade-off. The article states that it takes 20–30 seconds for o1 to answer, compared with under 3 seconds for GPT-4o. It also notes that the API cost for o1 is “many times higher” than standard models. OpenAI characterizes o1 as a model for “complex tasks,” not a replacement for GPT-4o.

Why later models still struggled

The pattern returned with newer releases. In December 2025, OpenAI released GPT-5.2. When asked the same question, the newest model again answered 2.

The article attributes this to tokenization and the model’s trade-offs. It says GPT-5.x uses a newer tokenization scheme named o200k_harmony, but “strawberry” still breaks into str + aw + berry. It also states that the chain-of-thought reasoning used by o1 was not carried over into GPT-5.x, because regular users typically do not want to wait 20–30 seconds for every answer.

Developments by April 2026 and comparisons with other systems

As of April 2026, the article says GPT-5.3 and GPT-5.4—the latest models—partly address the issue. It also claims that other systems, including Claude, Gemini, Grok, and Perplexity, “rarely” answer incorrectly, attributing this to different tokenization approaches.

It adds that o1-preview and o1-mini were deprecated since April 2025, but the “Strawberry” naming joke has continued.

Why the name “Strawberry” mattered

The article concludes that the most notable aspect may be the choice of name itself: a reminder of a specific limitation the team aimed to overcome. In an industry where new products are often marketed with broad, promotional language, “Strawberry” is presented as an unusually direct and candid label for a technical challenge.

By Thế Duyệt